EP3671576A1 - Procédé et dispositif de détermination des segments dans des données de séries temporelles reçues d'un composant de système - Google Patents

Procédé et dispositif de détermination des segments dans des données de séries temporelles reçues d'un composant de système Download PDF

Info

Publication number
EP3671576A1
EP3671576A1 EP18214350.3A EP18214350A EP3671576A1 EP 3671576 A1 EP3671576 A1 EP 3671576A1 EP 18214350 A EP18214350 A EP 18214350A EP 3671576 A1 EP3671576 A1 EP 3671576A1
Authority
EP
European Patent Office
Prior art keywords
time series
series data
sections
data
time
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Withdrawn
Application number
EP18214350.3A
Other languages
German (de)
English (en)
Inventor
Markus Michael Geipel
Stephan Merk
Sebastian MITTELSTÄDT
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Siemens AG
Original Assignee
Siemens AG
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Siemens AG filed Critical Siemens AG
Priority to EP18214350.3A priority Critical patent/EP3671576A1/fr
Priority to PCT/EP2019/085679 priority patent/WO2020127285A1/fr
Publication of EP3671576A1 publication Critical patent/EP3671576A1/fr
Withdrawn legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks

Definitions

  • the present invention relates to a method for determining segments in received time series data of a system component.
  • the invention further relates to a device for determining segments in received time series data of a system component.
  • the availability of marked data is one of the most important prerequisites for a successful machine learning process.
  • Training data is required to be able to carry out machine learning.
  • the training data for example of a machine, is provided with a label by an expert for the machine and made available.
  • the training data marked with labels shows when, for example, a machine works well, when the machine works poorly and when a certain incident (anomaly) has occurred.
  • the data marked with a label can be used for machine learning, for example, to automatically show when the machine is running well or badly, or when a certain incident has occurred.
  • the process of viewing and labeling the data with labels is, on the one hand, time and cost intensive because this process is carried out manually.
  • the process is very susceptible to errors due to the manual execution.
  • the labeled data lacks accuracy, to enable error-free machine learning.
  • a label in the time series data must cover exactly the corresponding information, for example an anomaly that is interesting for machine learning, and must not contain any additional information that can no longer be assigned to the anomaly. This would lead to incorrect teaching and errors in the subsequent productive process.
  • experts in their areas tend to label the data with the knowledge of other experts when labeling the data. The data is therefore not marked with the necessary care and may include too much information or too little information in the areas marked by a label. In most cases, this training data is then insufficient for machine learning and cannot be used.
  • the invention relates to a method for determining segments in received time series data of a system component with the step of receiving time series data from at least one system component over a specific period of time.
  • the time series data received by a system component are broken down into separate time series sections.
  • features are determined based on feature definitions.
  • the time series sections with the same characteristics are each assigned to a cluster.
  • the cluster is formed on the basis of a predefined probalistic description model.
  • a hidden Markow model is applied to the characteristics in the time series sections.
  • a state of the hidden Markow model corresponds to the clusters formed.
  • those time series sections are selected which have the same state of the hidden Markow model over the time period of the time series data.
  • those selected time series sections are assigned to each segment, which are consecutive in time.
  • the invention also provides a device with the features specified in claim 8.
  • the device for determining segments in received time series data of a system component comprises a receiving unit for receiving time series data from at least one system component over a specific period of time. Furthermore, the device comprises a decomposition unit for decomposing the received time series data into separate time series segments. In addition, the device comprises a determination unit for determining features in the time series data of the separate time series sections based on feature definitions. Furthermore, the device comprises an assignment unit for assigning the time series sections with the same features to a respective cluster, which is formed on the basis of a predefined probabilistic description model. There is also an application unit for applying a hidden Markow model on the features provided in the time series sections. A state of the hidden Markow model corresponds to the clusters formed. Furthermore, the device has a selection unit for selecting those time series sections which have the same state of the hidden Markow model over the time period of the time series data. In addition, there is an assignment unit for assigning to each segment those selected time series sections which are consecutive in time.
  • the respective unit can be implemented in terms of hardware and / or software.
  • the respective unit can be designed as a device or as part of a device, for example as a computer or as a microprocessor.
  • the respective unit can be designed as a computer program product, as a function, as a routine, as part of a program code or as an executable object.
  • the computer program product can be distributed across different and distributed computers, for example a computer network in the cloud.
  • the invention also provides a computer program product with the features specified in patent claim 9.
  • the invention further provides, according to a further aspect, a provision device for the computer program product with the features specified in claim 10.
  • Time series data is to be understood as recorded information / data, for example stored on a data carrier, which results from an observation of a system and / or a system component of a system by means of a sensor system which is in communication with the system and / or the system component. result over several successive points in time.
  • a segment is to be understood as a section or part of the time series data that has a constant dynamic of the time series data. Segments are equal to each other if they have the same dynamics in the time series data.
  • the segments receive, for example, coding, numbering and / or color assignment according to the dynamics in the time series data. Assignment of the segments with different dynamics to one another is not limited to the assignment features mentioned. Furthermore, further assignment features can be used to illustrate different dynamics in the time series data.
  • system generally refers to a set of elements that are related to one another and interact with one another in such a way that they can be viewed as a task, meaning or purpose-bound unit.
  • a system component is a single one of the components that make up the system.
  • the time period specifies the time over which the time series data of a system component are recorded or received.
  • the time period can comprise, for example, a range of microseconds, milliseconds, seconds, hours, days, weeks.
  • the time series data is cut into time series sections for better processing.
  • a segment can comprise several time series segments.
  • an expert can define which size of the time series sections is appropriate for the corresponding system component in order to record usable data for determining segments.
  • a cluster is a set of objects, for example data, with the same properties.
  • the data can be assigned to a cluster according to a normal state of a system component with the same dynamics and data can be assigned to a cluster different from the cluster with the same dynamics according to an anomaly in the system component.
  • the present invention is based on the knowledge that the generation of large amounts of example data, which are necessary for machine learning, represents a large expenditure of time and work for an expert.
  • larger amounts of data are processed in shorter time periods and can be labeled with a label after the assignment to segments.
  • the process of assigning to segments and labeling the segments with a label is quicker and easier to perform.
  • an expert no longer has to manually select the start time and the end time of the data for labeling with a label. Rather, the present invention predetermines the start and end time based on the specific segments. As a result, time series data can be labeled more quickly and at a lower cost.
  • the resulting labels are mutually consistent in the sense that segments with the same dynamics in the time series data cannot be labeled with a label for normal functioning of the system component and with a label for abnormal functioning of the system component at the same time.
  • the labels are placed more precisely and the time series data marked with the label are more precise and lead to improved machine learning.
  • the advantages mentioned above lead to improved training data for anomaly detection models and the quality of the machine learning model is improved.
  • an anomaly detection can be carried out in a system. It can be assumed that these anomalies have not yet appeared in the system. It can thus be determined which semantics are behind which data and these can be labeled with a label, for example for normal operation. In addition, a warning can be issued if a current operating state does not correspond to normal operation, regardless of which operating state is present and whether this operating state is known.
  • the assignment of the time series sections with the same features to a cluster is repeated for the segments in one step.
  • a distribution of the data of the time series data can thus advantageously be checked again. If it can be assumed that data changes the membership of a cluster, the distribution across the clusters can be re-estimated and the hidden Markov model can be calculated further. If a first segmentation has taken place, this process can advantageously be repeated repeatedly in an advantageous manner.
  • the data of the time series data in a time series section was assigned to a segment. The data can be reassigned for a review to check the distribution become.
  • the distribution of the data is recalculated and the allocation over time is recalculated using the hidden Markov model. This process continues until the assignment of the data to a segment no longer changes.
  • the transition matrix of the hidden Markow model is determined empirically.
  • the transition matrix is advantageously based on experience gained about the system component from operation.
  • the transition matrix of the hidden Markow model is based on the system data of an assistance system.
  • the described assistance system can represent a further computing unit connected to the device according to the present invention via a communication system.
  • the assistant system can further specify the transition matrix of the hidden Markow model based on learned models and thus adapt it to changing dynamics in the time series data of the system component.
  • these learned models can be based on the knowledge of an expert for the system component.
  • an expert can specify the transition matrix of the hidden Markow model.
  • the assistance system can advantageously assess how often the time series data of the system component change its dynamics. If the received time series data includes a segment change that is too frequent and that is atypical for the system component based on empirical values, the assistance system can intervene. It is also advantageous if, for example, the system component is too stable according to the time series data, although the segments are more dynamic and more often a segment change occurs in the time series data that the behavior can be adjusted. This behavior can be controlled by specifying a transition matrix.
  • the temporal dimension for the selection of the clusters is advantageously taken into account by means of the hidden Markow model.
  • the time series data are viewed over the temporal dimension and frequent changes in cluster membership are taken into account and smoothed.
  • the Hidden Markow Model is an optimization process with an error function, which can be used to assign data according to selected criteria. Additional models can be used as an optimization function as long as the models take into account frequent changes in cluster membership.
  • the size of the time series sections can be configured.
  • the time series sections can advantageously be adapted to the respective received time series data of different system components. Dynamics with high frequency as well as with low frequency can thus be analyzed in the time series data.
  • the feature definitions include the level of the time series data, a noise of the time series data, the gradient of the time series data or the standard deviation.
  • the time series data received can thus be divided into sections with the same dynamics on the basis of the feature definitions.
  • the same dynamics can be determined in the time series data and assigned to a cluster. For example, by the method according to the invention in certain areas the time series data found a high variance, which have a significantly higher level. These areas have the same dynamics and can be assigned to a cluster.
  • the probalistic description model is a Gaussian mixing model or a mixing model based on a student t-distribution.
  • the Gaussian distribution can advantageously be transformed again.
  • the assignment of the time series sections is not limited to the Gaussian assignment to clusters. Furthermore, further parametric distributions can be used and / or these can be used in a common assignment to clusters. Each cluster should have a probalistic interpretation so that the probability of a point in the feature space of a particular cluster can be calculated.
  • the mixing model based on the student t-distribution is more robust compared to outliers in the time series data, which can lead to a deformation in the Gaussian mixing model.
  • the mixing model based on the student t-distribution corrects less and is therefore more robust.
  • the feature definitions include the variance, the mean, the median, or the logarithm of the variance.
  • the feature definitions mentioned here do not constitute a conclusive list.
  • other features that describe a distribution of the time series data can be used.
  • the time series data of the system component is prepared for semantic assignment.
  • the semantic assignment is relevant for machine learning and improves the efficiency and quality of machine learning.
  • An assigned label for a segment thus represents the semantic evaluation of this section.
  • Figure 1 shows schematically the sequence of the method for determining segments 15, 16, 17, 18 in received time series data 10 of a system component 1.
  • a system component 1 is provided to provide time series data 10.
  • the time series data 10 are received over a specific time period t in a first step S1.
  • the received time series data 10 is broken down S2 into successive, for example overlapping, time series sections 11-i of configurable size.
  • Numeral 11-i refers to all time series segments as a whole.
  • Reference numerals 11-1 designate the first time series section of the time series data 10 with respect to the time period t, where t represents the maximum temporal value of the time series data 10.
  • An assistance system 2 provides feature definitions.
  • features are determined for each signal in the time series data 10 of the time series sections 11-i S3.
  • the feature definitions characterize the time series sections 11-i which belong to the same system state of the system component 1. For example, this can include the statistical measurement of the variance and the mean value for a sensor A and the gradient and the variance for a sensor B of the system component 1.
  • the time series sections 11-i with the same characteristics are each assigned to a cluster 12, 13, 14 S4.
  • the cluster 12, 13, 14 is formed on the basis of a predefined probalistic description model 3.
  • the probalisitic description model 3 can be a Gaussian mixing model in which each cluster 12, 13, 14 corresponds to a Gaussian distribution.
  • the Gaussian distribution can be defined, for example, by the mean vector and variance-covariance matrix.
  • the probalistic description model 3 can have a mixing model based on a student t-distribution. Can also further mixing models not listed can be used.
  • a hidden Markow model is applied to the features 4 in the time series sections 11-i. Each of the states of the hidden Markow model corresponds to a cluster formed in the probalistic description model 3.
  • the transition matrix can be estimated empirically or determined on the basis of knowledge about the system component 1 by an assistance system 2. Those time series sections 11-i are selected S6 which have the same state of the hidden Markow model over the time duration of the time series data 10.
  • the most likely sequence of states for time series data 10 can be calculated, for example, by using the Viterbi algorithm.
  • the most probable sequence of states describes the segmentation of the time series data 10. Those selected time series sections 11-i are each assigned to a segment 15, 16, 17, 18, which are S7 successively in time.
  • one possible embodiment provides for refining the cluster description on the basis of the segments found. This is done by repeatedly assigning S8 the time series sections 11-i with the same features to a cluster. This procedure follows the pattern of maximizing expectations, which is also used in classic K-means clustering.
  • time series data 10 are shown schematically.
  • the time series data 10 shown have time series sections 11-i with features 4.
  • the time series sections 11-i with the same features 4, for example the same dynamics, are assigned to clusters 12, 13, 14.
  • Time series sections 11-i with the same dynamics are assigned to the same cluster 12, 13, 14.
  • the time series sections 11-i in cluster 12 have the same dynamics.
  • the time series sections 11-i in clusters 13 and 14 have different dynamics in comparison each other and cluster 12 up.
  • Those time series sections 11-i are selected which have the same state of the hidden Markow model over the time period t of the time series data 10.
  • the selected time series sections 11-i, which are consecutive in time are assigned to a segment 15, 16, 17, 18.
  • each cluster 12, 13, 14 has a different dynamic.
  • the same dynamics of the two clusters 12 are separated by the dynamics of the cluster 13.
  • each selected time series section 11-i, or each cluster 12, 13, 14 is assigned to a segment 15, 16, 17, 18.
  • the method of the present invention assigns each date of the time series data 10 to a cluster 12, 13, 14, for example the cluster 12.
  • Different segments 15 and 17 correspond to the cluster 12, which represent different time series sections 11-i in the time series data 10 with the same dynamics, as predefined for example by the feature definitions, but do not lie together in time ( Fig. 2 ).
  • the same segments, for example segment 15, correspond to the cluster 12, which represent different time series sections 11-i in the time series data 10 with the same dynamics, as predefined for example by the feature definition, and which are consecutive in time ( Fig. 3 ).
  • the segments determined using the method and the device 20 can be used for labeling with labels by an expert in two different applications.
  • a user of the method, and the device that directly labels segments 15, 16, 17, 18 with labels instead of manually creating the labels with a new start and end time. This can be done, for example, in a desktop application or a web application by selecting a specific segment 15, 16, 17, 18 by selecting and assigning a label.
  • all segments are also automatically assigned the same label, which correspond to the same cluster 12 and thus to the same dynamics of the system component according to the feature definition.
  • segments of the time series data 10 are already provided with labels. These were assigned, for example, by an expert in system component 1.
  • the method and the device for analyzing and checking the consistency of the labels assigned by the experts can be used. The consistency check can lead to feedback to the experts, for example to refine the selection of the labels on the segments and / or to refine the relevant feature definitions, to remove the inconsistency and to improve the quality of the marked time series data for machine learning.
  • Figure 3 shows schematically the sequence of the method for determining segments 15, 16, 17, 18 in received time series data 10 of a system component 1.
  • a system component 1 is provided to provide time series data 10.
  • the time series data 10 are broken down into separate time series sections 11-i in a step S2.
  • the number of time series sections 11-i into which the time series data 10 is broken down corresponds to a configuration parameter depending on the system or the system component.
  • the number of time series sections 11-i over the time period t of the time series data 10 can be selected such that a sufficiently high number of time series data 10 in the time series data 10 of time series sections 11-i for determining segments 15, 16, 17, 18 is present.
  • features 4 are calculated in a step S3 based on feature definitions.
  • the features 4 include the mean and the variance.
  • the mean value corresponds to the value 1.1 and the variance to the value 2.6 of the time series data 10 in the first time series section 11-1 of the time series data 10.
  • other features and / or combinations of features can also be calculated.
  • the received time series data 10 (raw data) were converted into a time series of features 4.
  • a probabilistic description model 3 can be applied to the time series of features 4.
  • the probabilistic description model 3 can have a Gaussian mixing model or a mixing model based on a student's t-distribution.
  • the time series sections 11-i are assigned the same features 4 to a cluster 12, 13, 14.
  • the cluster 12, 13, 14 is formed on the basis of a Gaussian mixing model.
  • Each cluster 12, 13, 14 found can be represented by a Gaussian normal distribution or a separate Gauss bell.
  • three different Gauss bells were identified and each Gauss bell can be assigned to a respective cluster.
  • a hidden Markow model can be applied to the series of features 4 assigned to clusters 12, 13, 14.
  • the hidden Markow model has a large number of discrete states which correspond to the number of clusters 12, 13, 14.
  • the hidden Markow model takes the time information into account and thus smoothes the time series data 10. It can be assumed here that there is no constant change between the individual clusters. For a change of state and thus Appropriate evidence is required in another cluster 12, 13, 14. For example, after the first time series section 11-1 assigned to cluster 12 with the value 1.1; 2.6, there is a change to cluster 13 (time series section 11-2 with value 1.3; 2.1).
  • the time series sections 11-i are designated in ascending order in the course of the time period t of the time series data 10.
  • the hidden markow model is used to reallocate the time series sections 11-i to corresponding clusters.
  • the time series section 11-2 is assigned to the cluster 13 according to the determined features 4.
  • the time series section 11-1 and the time series section 11-3 which lie before and after the time series section 11-2 are assigned to the cluster 12.
  • there is a cluster change except for the features 4 of the time series section 11-2, there is insufficient evidence for a cluster change.
  • the time series section 11-2 is reassigned to cluster 12 by the hidden Markow model. An assignment is thus advantageously determined which avoids a cluster change but is at the same time plausible.
  • the time series sections are reallocated over all time series sections 11-i of the time series of the features, which increases the consistency of the cluster changes.
  • the selected time series sections 11-1 which lie consecutively in time, are each assigned to a segment 15, 16, 17, 18.
  • the selected time series sections 11-1, 11-2, 11-3, 11-4 have the same state of the hidden Markow model over the time period t and are located successively.
  • the selected time series sections 11-1, 11-2, 11-3, 11-4 are thus assigned to segment 15.
  • the selected time series sections 11-1, 11-2, 11-3, 11-4 are assigned to the same cluster 12 as the time series sections 11-9, 11-10.
  • the time series sections 11-9, 11-10 are assigned to the segment 17.
  • time series sections 11-5, 11-6, 11-7, 11-8 between the cluster 12 comprising the time series sections 11-1, 11-2, 11-3, 11-4 and the cluster 12 comprising the time series sections 11-9 , 11-10 features 4 different from the features in cluster 12 were determined.
  • the time series sections 11-5, 11-6, 11-7, 11-8 are assigned to the cluster 13.
  • segments with a label can be identified by an assistance system 2.
  • segments comprising the same cluster membership can be provided with the same label by an assistance system 2.
  • FIG 4 shows schematically different time series data 10 from six different system components B1 to B6.
  • the time series data 10 were received by the system components B1 to B6 over a period of time t.
  • both the data for a normal state NZ of the system component in operation and for a poor state of the system component in operation are to be determined.
  • an anomaly detection should take place via the time series data 10.
  • the corresponding areas (normal condition NZ, poor condition, abnormal condition AZ) must be identified by a label. These labels are the prerequisite for machine learning with training data.
  • One application of the invention is in the area of preventive maintenance for oil and gas technology with artificial intelligence.
  • the invention can be applied to electric submersible pumps.
  • sample data or training data for teaching or training the artificial intelligence model are required.
  • the operator of the submersible pumps should provide data marked with labels for creating the training data.
  • the labels represent time ranges that are specific for different operating states for a submersible pump. This data marked with labels is not available to the pump operators and often has to be created first, which is time-consuming and costly.
  • time series data 10 from electric submersible pumps can be segmented by means of the method and the device 20 according to the invention.
  • the segments 15, 16, 17 correspond to different operating states.
  • the expert for example the oil and gas engineer, no longer has to manually identify the individual operating states of the submersible pump in the time series data 10, but can verify the respective segments relating to the corresponding operating states of the submersible pump and assign appropriate labels. This advantageously leads to a reduction in manual effort and increases the quality of the labels. With a high number of high quality Labels, the machine learning of the artificial intelligence models can be carried out with less effort.
  • FIG. 5 shows schematically different time series data 10 from four different system components B1, B3, B4 and B6.
  • Time series data 10 of an anomaly for example the abnormal state AZ of a blockage of the submersible pump, are also shown.
  • the time series data 10 are assigned the labels for normal state NZ and abnormal state AZ of the submersible pump.
  • the time series data 10 In the areas marked with the label for normal state NZ, the time series data 10 have the same dynamics.
  • the time series data 10 In the area marked with the abnormal condition AZ label, the time series data 10 have a dynamic different from the dynamic in the normal NZ condition.
  • machine learning of the artificial intelligence models can take place.
  • FIG 6 are the time series data 10 according to the Figure 4 shown.
  • the time series data 10 are assigned to the clusters 12, 13, 14 in accordance with their dynamics in accordance with the feature definitions.
  • a hidden markow model is applied to the features in the time series sections 11-i, a state of the hidden markow model corresponding to the cluster 12, 13, 14 formed. From the time series sections 11-i, those are selected which have the same state of the hidden Markow model over the time period of the time series data.
  • the selected time series sections which are consecutive in time are each assigned to a segment 15, 16, 17, 18.
  • the segments 15, 16, 17, 18 which correspond to a same cluster 12, 13, 14 can be identified by an assistance system 2 with the same label.
  • FIG 7 are the time series data 10 according to the Figure 6 shown. Furthermore, the time series data 10 of the Figure 7 a segment 19 determined by an expert, which with a Label was marked by the experts. In this regard, the indicator of the time series shows an inconsistency.
  • Each segment 15, 16, 17, 18 determined by the method according to the invention has a different dynamic and is therefore to be labeled with different labels.
  • the expert applied a label across all three segments 15, 16, 17, 18. This means that either the characteristics used do not correspond to the reasoning of the expert of the system component or the expert has incorrectly labeled them. According to the expert, segment 16, for example, can be regarded as not relevant for machine learning. This can lead to incorrect machine learning.
  • each segment 15, 16, 17, 18 corresponds to a cluster 12, 13, 14.
  • the corresponding cluster 12, 13, 14 is labeled with a label, and all clusters connected to it 12, 13, 14.
  • time ranges in the time series data 10 that are in the same cluster are not assigned different labels.
  • the labeling with labels is thus consistent with the dynamics within the time series data 10. In this way, labeling of a segment 18 with a label via segments of different dynamics can advantageously be avoided and the consistency and quality of the labels can be increased.
  • Reference numeral 20 denotes a device for determining segments in received time series data 10 of a system component 1.
  • the device comprises a receiving unit 21 for receiving time series data 10 from at least one system component 1 over a specific time period t.
  • the received time series data 10 is divided into separate time series sections by means of a decomposition unit 22.
  • a number of time series sections 11-i can be assigned to a segment 15, 16, 17, 18.
  • the time series data 10 must be used to determine which temporal resolution of the Time series section for an analysis of the dynamics in the time series data 10 is useful. The temporal resolution depends on the system component 1 or the system and the dynamics to be analyzed, for example caused by anomalies.
  • the characteristics can include, for example, the average level (mean) and the range of variation (variance).
  • an allocation unit 24 assigns the time series sections with the same features to a cluster 12, 13, 14 on the basis of a probalistic description model. All values of the time series data 10 are thus assigned to a respective group (cluster).
  • an application unit 25 applies a hidden markow model to the features 4 in the time series sections 11-1, a state of the hidden markow model corresponding to the cluster 12, 13, 14 formed.
  • the Hidden Markow Model analyzes how high the probability is that the cluster is constantly changing. It is assumed that a system component has a certain dynamic and no constantly changing, so that each time series section 11-i has different dynamics.
  • the hidden Markow model can be used to determine, for example in the case of the present of the characteristic values of the mean value and the variance in a time series section 11-i, that one group has a low level and another group has a low variance. In addition, other groups can have a high variance. It is thus established that the system component 1 has different dynamics, these being characterized by the mean and the variance. These characteristics are analyzed over time for a change in their dynamics. This means that the dynamics have a high probability that they will change and must therefore be assigned to another cluster.
  • the selection unit 26 selects those time series sections 11-i which have the same state of the hidden Markow model over the time period t of the time series data 10.
  • those selected time series sections 11-i are each assigned to a segment 15, 16, 17, 18, which are consecutive in time.
  • the time series data 10 received can be identified more efficiently and more precisely with a label by an expert.
  • FIG. 9 A provision device 40 is shown.
  • the provisioning device 40 stores or provides the computer chart product 30.
  • Computer program product 30 is configured, when executed, to perform the method according to the present invention.
  • the segments in the received time series data 10 of a system component 1 are determined with the computer program product 30.
  • the invention relates to a method for determining segments in received time series data of a system component with the step of receiving time series data from at least one system component over a specific period of time.
  • the time series data received by a system component are broken down into separate time series sections.
  • features are determined based on feature definitions.
  • the time series sections with the same characteristics are each assigned to a cluster.
  • the cluster is formed on the basis of a predefined probalistic description model.
  • a hidden Markow model is applied to the characteristics in the time series sections.
  • a state of the hidden Markow model corresponds to the clusters formed.
  • those time series sections are selected which have the same state of the hidden Markow model over the time period of the time series data.
  • those selected time series sections are assigned to each segment, which are consecutive in time.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • General Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computational Mathematics (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Optimization (AREA)
  • Mathematical Analysis (AREA)
  • Algebra (AREA)
  • Probability & Statistics with Applications (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
EP18214350.3A 2018-12-20 2018-12-20 Procédé et dispositif de détermination des segments dans des données de séries temporelles reçues d'un composant de système Withdrawn EP3671576A1 (fr)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP18214350.3A EP3671576A1 (fr) 2018-12-20 2018-12-20 Procédé et dispositif de détermination des segments dans des données de séries temporelles reçues d'un composant de système
PCT/EP2019/085679 WO2020127285A1 (fr) 2018-12-20 2019-12-17 Procédé et dispositif de détermination de segments dans des données de séries temporelles reçues d'un composant de système

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
EP18214350.3A EP3671576A1 (fr) 2018-12-20 2018-12-20 Procédé et dispositif de détermination des segments dans des données de séries temporelles reçues d'un composant de système

Publications (1)

Publication Number Publication Date
EP3671576A1 true EP3671576A1 (fr) 2020-06-24

Family

ID=64746314

Family Applications (1)

Application Number Title Priority Date Filing Date
EP18214350.3A Withdrawn EP3671576A1 (fr) 2018-12-20 2018-12-20 Procédé et dispositif de détermination des segments dans des données de séries temporelles reçues d'un composant de système

Country Status (2)

Country Link
EP (1) EP3671576A1 (fr)
WO (1) WO2020127285A1 (fr)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3992739A1 (fr) 2020-10-29 2022-05-04 Siemens Aktiengesellschaft Génération automatique de données d'apprentissage d'une série temporelle de données de capteur
CN115034337A (zh) * 2022-08-10 2022-09-09 江西科骏实业有限公司 一种轨道交通车辆中牵引电机状态辨识方法及装置、介质

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
JENS KOHLMORGEN ET AL: "A Dynamic HMM for On-line Segmentation of Sequential Data", PROCEEDINGS OF THE 14TH INTERNATIONAL CONFERENCE ON NEURAL INFORMATION PROCESSING SYSTEMS: NATURAL AND SYNTHETIC, 3 December 2001 (2001-12-03), pages 793 - 800, XP055596809 *
YUNZHAO JIA ET AL: "Symbolic Important Point Perceptually and Hidden Markov Model Based Hydraulic Pump Fault Diagnosis Method", SENSORS, vol. 18, no. 12, 17 December 2018 (2018-12-17), pages 4460, XP055596878, DOI: 10.3390/s18124460 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3992739A1 (fr) 2020-10-29 2022-05-04 Siemens Aktiengesellschaft Génération automatique de données d'apprentissage d'une série temporelle de données de capteur
WO2022090275A1 (fr) 2020-10-29 2022-05-05 Siemens Aktiengesellschaft Génération automatique de données d'apprentissage d'une série chronologique de données de capteur
CN115034337A (zh) * 2022-08-10 2022-09-09 江西科骏实业有限公司 一种轨道交通车辆中牵引电机状态辨识方法及装置、介质
CN115034337B (zh) * 2022-08-10 2022-11-01 江西科骏实业有限公司 一种轨道交通车辆中牵引电机状态辨识方法及装置、介质

Also Published As

Publication number Publication date
WO2020127285A1 (fr) 2020-06-25

Similar Documents

Publication Publication Date Title
DE102019107363B4 (de) Verfahren und System zum Bestimmen einer Eigenschaft einer Maschine, insbesondere einer Werkzeugmaschine, ohne messtechnisches Erfassen der Eigenschaft sowie Verfahren zum Bestimmen eines voraussichtlichen Qualitätszustands eines mit einer Maschine gefertigten Bauteils
EP3841732B1 (fr) Procédé, système informatique et programme informatique destiné au traitement automatique des identificateurs de données
DE102019124018A1 (de) Verfahren zum Optimieren von Tests von Regelsystemen für automatisierte Fahrdynamiksysteme
DE102016117585A1 (de) Verfahren und Vorrichtung zur Bestimmung von Schritten für eine multi-Zufallsvariablen Batch-Steuerungs-Analyse
EP3671576A1 (fr) Procédé et dispositif de détermination des segments dans des données de séries temporelles reçues d'un composant de système
DE102020207449B4 (de) Verfahren, Computerprogramm und Vorrichtung zum Verarbeiten von Signalen
EP3951529A1 (fr) Dispositif et procédé de surveillance pour la détection d'anomalies
DE102021204550A1 (de) Verfahren zum Erzeugen wenigstens eines Datensatzes zum Trainieren eines Algorithmus maschinellen Lernens
EP3407273A1 (fr) Procédé et dispositif de détermination de l'état anormal d'un système
EP3454154A1 (fr) Détection automatique des dépendances statistiques entre les messages du processus
DE102012112351A1 (de) Verfahren zur Aufnahme von Systemzuständen mit einer Vorrichtung sowie entsprechend ausgestaltete Vorrichtung
EP3716578B1 (fr) Procédé et dispositif de commande d'un appareil technique à l'aide d'un modèle optimal
EP3929554A1 (fr) Détection améliorée d'erreurs dans les machines au moyen de l'ia
DE102020201183A1 (de) Verfahren und Vorrichtung zur Simulation eines technischen Systems
EP2990941A1 (fr) Procédé assisté par ordinateur de production d'un code de programme de dispositif de commande et environnement correspondant de gestion de messages
EP3705993A1 (fr) Système et procédé de détection et d'identification des n uds de calcul dans un réseau
DE102018212801A1 (de) Diagnose komplexer Systeme
EP3483677B1 (fr) Procédé d'incorporation dans une interface utilisateur d'au moins un dispositif de surveillance et / ou de commande, comportant une interface de données, d'un système de surveillance et / ou d'automatisation
EP4341758A1 (fr) Procédé, appareil, programme informatique et support de stockage lisible par ordinateur pour déterminer des véhicules défectueux
EP4300234A1 (fr) Dispositif de surveillance destiné à surveiller l'état d'une machine
DE102023201929A1 (de) Verfahren und Vorrichtung zum Bereitstellen von Konfigurationsvorschlägen für eine Konfiguration eines modularen Produktionssystems
DE202023104442U1 (de) Auf Software-Evolution basierendes Aspekt-Mining-System für Data Mining
DE112021008086T5 (de) Zustandserfassungssystem, zustandserfassungsverfahren und zustandserfassungsprogramm
WO2023202778A1 (fr) Procédé et système mis en œuvre par ordinateur pour détecter des anomalies pendant le fonctionnement d'un dispositif technique
WO2023209029A1 (fr) Procédé et système mis en œuvre par ordinateur pour le fonctionnement d'un dispositif technique à l'aide d'un modèle basé sur un apprentissage fédéré

Legal Events

Date Code Title Description
PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE APPLICATION HAS BEEN PUBLISHED

AK Designated contracting states

Kind code of ref document: A1

Designated state(s): AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR

AX Request for extension of the european patent

Extension state: BA ME

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE APPLICATION IS DEEMED TO BE WITHDRAWN

18D Application deemed to be withdrawn

Effective date: 20210112